kha-en-nllb-v0.4

**MWire Labs | Internal Use Only **

Results (goldtest_new.xlsx — 500 pairs)

Model	kha→en	en→kha
stage0.1	17.25	21.78
stage1	17.94	21.80
v0.3	32.31	35.83
v0.4 (this)	32.97	36.80

Training

• Base: Badnyal/kha-en-nllb-v0.3
• Epochs: 3 | LR: 5e-6 cosine | bf16 | batch 32
• Val loss: 0.3270 (vs 0.5231 in v0.3)

Language Token

kha_Latn (ID: 256204) — manually added, not in base NLLB vocab.

Inference

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("Badnyal/kha-en-nllb-v0.4")
model = AutoModelForSeq2SeqLM.from_pretrained("Badnyal/kha-en-nllb-v0.4")
model.eval()

def translate(text, src_lang, tgt_lang):
    tokenizer.src_lang = src_lang
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
    forced_bos_id = tokenizer.convert_tokens_to_ids(tgt_lang)
    outputs = model.generate(**inputs, forced_bos_token_id=forced_bos_id, max_new_tokens=128, num_beams=5)
    decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)
    for tag in ["eng_Latn", "kha_Latn"]:
        if decoded.startswith(tag):
            decoded = decoded[len(tag):].strip()
    return decoded

# kha -> en
print(translate("Ka jingthmu jong ka sorkar jylla", "kha_Latn", "eng_Latn"))
# en -> kha
print(translate("The government is working to develop the state.", "eng_Latn", "kha_Latn"))